Granted Patent B2

US 12,536,123 · App. 18/383,745 · Granted Jan 27, 2026

Inventors: Manish K. Shah (Austin, TX); Ram Sivaramakrishnan (San Jose, CA); Gregory Frederick Grohoski (Bee Cave, TX); Raghu Prabhakar (San Jose, CA)

Assignee: SambaNova Systems, Inc.

G06F13/42

Abstract

Claims (45)

1 . A reconfigurable processing unit comprising:

a first internal network and a second internal network, the first internal network and the second internal network having different protocols;

an interface to an external network, the external network having a protocol different than either the first internal network or the second internal network;

a first configurable unit coupled to the first internal network;

a second configurable unit coupled to both the first internal network and the second internal network; and

a third configurable unit coupled to both the second internal network and the interface to the external network, and configured to receive a payload containing a transaction type identifier and an identifier of the second configurable unit through the interface to the external network, and send a first packet including the transaction type identifier to the second configurable unit over the second internal network;

the second configurable unit configured to, in response to the transaction type identifier being a particular type, increment a counter; and

the second configurable unit further configured to determine that the first configurable unit is in an execution state, and in response while the counter is non-zero, send a token over the first internal network to the first configurable unit and decrement the counter.

2 . The reconfigurable processing unit of claim 1 , the first configurable unit configured to increment a buffer size counter in response to receiving the token.

3 . The reconfigurable processing unit of claim 2 , the first configurable unit further configured to determine whether the buffer size counter has a value greater than zero, and in response to the buffer size counter having a value greater than zero, decrement the buffer size counter and send data over the first internal network to the second configurable unit;

the second configurable unit further configured to, in response to receiving the data from the first configurable unit, send a second packet including the data, an identifier for a target reconfigurable processing unit, and an identifier for a target configurable unit of the target reconfigurable processing unit to the third configurable unit over the second internal network; and

the third configurable unit further configured to, in response to receiving the second packet from the second configurable unit, determine an address on the external network for the target reconfigurable processing unit based on the identifier of the target configurable unit, synthesize a second payload that includes a stream write request, the data, and an identifier of the reconfigurable processing unit, and send the second payload to the address over the external network.

4 . The reconfigurable processing unit of claim 1 , the second configurable unit further configured to receive a third packet on the second internal network and queue the third packet in a FIFO of the second configurable unit;

wherein the first packet is not queued in the FIFO.

5 . The reconfigurable processing unit of claim 1 , the second configurable unit further configured to:

(a) receive a request to send a fourth packet including the particular type of transaction type over the second internal network;

(b) increment a counter in response to receiving the request;

(c) send the fourth packet including the particular type of transaction over the second internal network;

(d) decrement the counter in response to sending the fourth packet; and

(e) checking a value of the counter and in response to the value of the counter being greater than zero, repeating (c), (d), and (e) until the value of the counter is zero.

6 . A method for peer-to-peer communication between a first reconfigurable processing unit (RPU) and a second RPU, the first RPU having a second configurable unit, a first configurable unit coupled to the second configurable unit by a first internal network, and a third configurable unit coupled to the second configurable unit by a second internal network, the method comprising:

receiving a payload containing a transaction type identifier and an identifier of the second configurable unit over an external network at the third configurable unit;

sending a first packet including the transaction type identifier to the second configurable unit over the second internal network;

incrementing a counter in the second configurable unit in response to the transaction type identifier being a particular type;

determining, by the second configurable unit, that the first configurable unit is in an execution state; and

sending a token over the first internal network to the first configurable unit and decrementing the counter in response to said determination while the counter is non-zero.

7 . The method of claim 6 , further comprising incrementing a buffer size counter in the first configurable unit in response to receiving the token.

8 . The method of claim 7 , further comprising:

determining that the buffer size counter in the first configurable unit has a value greater than zero;

decrementing the buffer size counter;

sending data over the first internal network from the first configurable unit to the second configurable unit;

sending a second packet including the data, an identifier for the second RPU, and an identifier for a target configurable unit of the second RPU from the second configurable unit to the third configurable unit over the second internal network;

looking up an address on the external network for the second RPU at the third configurable unit based on the identifier of the second RPU;

synthesizing a second payload that includes a stream write request, the data, and an identifier of the second RPU; and

sending the second payload from the third configurable unit to the address over the external network.

9 . The method of claim 6 , further comprising:

receiving a third packet on the second internal network at the second configurable unit; and

queuing the third packet in a FIFO of the second configurable unit;

wherein the first packet is not queued in the FIFO.

10 . The method of claim 6 , further comprising:

(a) receiving a request to send a fourth packet including the particular type of transaction type over the second internal network at the second configurable unit;

(b) incrementing a counter in the second configurable unit in response to receiving the request;

(c) sending the fourth packet including the particular type of transaction over the second internal network from the second configurable unit;

(d) decrementing the counter in the second configurable unit in response to sending the fourth packet; and

(e) checking a value of the counter in the second configurable unit, and in response to the value of the counter being greater than zero, repeating (c), (d), and (e) until the value of the counter is zero.

Assignments (2)

Continuity (5)

References Cited (59)